Translation Invariant Feature Extraction Research Articles

Abnormal human activity recognition system based on R-transform and independent component features for elderly healthcare

This article presents a vision-based human activity recognition (HAR) system to recognize abnormal activities among a set of daily-life activities of elderly people. A novel method is proposed for feature extraction by the combination of R-transform, principal component analysis (PCA) and independent component analysis (ICA). R-transform is used for periodic, scale, and translation invariant feature extraction. PCA is used for dimensions reduction and ICA is used for independent features extraction. The features are processed by k-means to get observation sequences which are used by hidden Markov model for HAR. The dataset consisting of six abnormal activities (backward fall, forward fall, chest pain, headache, vomit, and faint) and three normal activities (walk, raise one hand, and raise both hands) from the view angles (90°, −90°, 45°, −45°, and 0°) is produced. Experimental results show average recognition rate of 86.86% for abnormal and normal activities from different view angles by our proposed system using (R-transform + PCA + ICA) as compared to R-transform, PCA, and ICA methods.

Rotation and Translation Invariant Feature Extraction Using Angular Projection in Frequency Domain

Rotation and Translation Invariant Feature Extraction Using Angular Projection in Frequency Domain

위치이동에 무관한 웨이블릿 변환을 이용한 패턴인식

웨이블릿 변환(Wavelet Transform)은 공간-주파수 영역에서 신호의 국소특성을 효율적으로 구현할 수 있다 하지만, 웨이블릿 변환을 패턴 인식을 위한 특징 추출에 적용할 경우, 입력 신호의 위치 이동에 따라 추출된 특징 값이 변화하게 되어 인식률이 낮아지는 결함이 있다. 본 논문에서는 웨이블릿 변환을 패턴 인식에 적용할 경우 발생하는 입력 신호의 위치 이동에 따른 문제점을 보완하여 노이즈에 강인한 홍채인식 알고리즘을 제안한다. 실험을 통하여 제안한 알고리즘의 우수성을 보여 준다. Wavelet Transform can effectively represent the local characteristics of a signal in the space-frequency domain. However, the feature vector extracted using wavelet transform is not translation invariant. This paper describes a new feature extraction method using wavelet transform, which is translation-invariant. Based on this translation-invariant feature extraction, the iris recognition method, based on this feature extraction method, is robust to noises. Experimentally, we show that the proposed method produces super performance in iris recognition.